Learning Sentential Paraphrases from Bilingual Parallel Corpora for Text-to-Text Generation

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1168–1179, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Learning Sentential Paraphrases from Bilingual Parallel Corpora for Text-to-Text Generation Juri Ganitkevitch, Chris Callison-Burch, Courtney Napoles, and Benjamin Van Durme Center for Language and Speech Processing, and HLTCOE Johns Hopkins University Abstract Previous work has shown that high quality phrasal paraphrases can be extracted from bilingual parallel corpora. However, it is not clear whether bitexts are an appropriate re- source for extracting more sophisticated sen- tential paraphrases, which are more obviously learnable from monolingual parallel corpora. We extend bilingual paraphrase extraction to syntactic paraphrases and demonstrate its abil- ity to learn a variety of general paraphrastic transformations, including passivization, da- tive shift, and topicalization. We discuss how our model can be adapted to many text gener- ation tasks by augmenting its feature set, de- velopment data, and parameter estimation rou- tine. We illustrate this adaptation by using our paraphrase model for the task of sentence compression and achieve results competitive with state-of-the-art compression systems. 1 Introduction Paraphrases are alternative ways of expressing the same information (Culicover, 1968). Automatically generating and detecting paraphrases is a crucial as- pect of many NLP tasks. In multi-document sum- marization, paraphrase detection is used to collapse redundancies (Barzilay et al., 1999; Barzilay, 2003). Paraphrase generation can be used for query expan- sion in information retrieval and question answer- ing systems (McKeown, 1979; Anick and Tipirneni, 1999; Ravichandran and Hovy, 2002; Riezler et al., 2007). Paraphrases allow for more flexible matching of system output against human references for tasks like machine translation and automatic summariza- tion (Zhou et al., 2006; Kauchak and Barzilay, 2006; Madnani et al., 2007; Snover et al., 2010). Broadly, we can distinguish two forms of para- phrases: phrasal paraphrases denote a set of surface text forms with the same meaning: the committee’s second proposal the second proposal of the committee while syntactic paraphrases augment the surface forms by introducing nonterminals (or slots) that are annotated with syntactic constraints: the NP1’s NP2 the NP2 of the NP1 It is evident that the latter have a much higher poten- tial for generalization and for capturing interesting paraphrastic transformations. A variety of different types of corpora (and se- mantic equivalence cues) have been used to auto- matically induce paraphrase collections for English (Madnani and Dorr, 2010). Perhaps the most nat- ural type of corpus for this task is a monolingual parallel text, which allows sentential paraphrases to be extracted since the sentence pairs in such corpora are perfect paraphrases of each other (Barzilay and McKeown, 2001; Pang et al., 2003). While rich syn- tactic paraphrases have been learned from monolin- gual parallel corpora, they suffer from very limited data availability and thus have poor coverage. Other methods obtain paraphrases from raw monolingual text by relying on distributional simi- larity (Lin and Pantel, 2001; Bhagat and Ravichan- dran, 2008). While vast amounts of data are readily available for these approaches, the …

Content truncated to 500 words